19 Jul
2010
19 Jul
'10
6:30 p.m.
I'm reporting this performances problem we encountered today. We have a large expression operating on table of 51200000 elements; In 32bits mode, our proto code for NT² works perfectly and output a gprofile like: http://codepad.org/Uyyvkq4X rutniem performance sis somethign like 1.5s for the whoel computation. In 64bits on linus using gcc 4.4 or 4.3, we got a massive 8.6s of runtime (!!) and the profile is: http://codepad.org/6VgRsbyR As it seems the call to proto::default_ is not inlined (hence the 5.12M calls) and take 80% of the runtime. I don't thing anything is 32/64bits specific in proto so is it a bug in gcc ? I don't have code handy here but will do tomorrow.